In response to long-standing efficiency bottlenecks in the AI speech synthesis field, Apple recently collaborated with Tel Aviv University to publish an innovative research titled "Principled Coarse-Grained" (PCG). This technology successfully improved the speed of speech generation by about 40% while ensuring "zero loss" in audio quality, by changing the way AI verifies sound predictions.
Currently, most mainstream text-to-speech (TTS) models use a "autoregressive" mechanism, where sounds are predicted one by one, like stringing beads. However, this approach is extremely rigid in terms of results, and the model often forces corrections even for minimal perceptual differences between the predicted output and the pre-set data. This not only consumes a lot of computing power but also significantly slows down the generation speed.
Apple's research team proposed PCG technology, which broke this deadlock. The core logic of this technology lies in "seeking common ground while reserving differences": researchers found that many subtle differences in sound segments are almost identical in human perception. Therefore, PCG introduced the concept of "acoustic similarity groups," upgrading the traditional "exact point verification" to "range verification." As long as the AI-generated prediction falls within a reasonable acoustic range, the system will directly accept it.
In practical tests, PCG performed remarkably. Even when 91.4% of the speech segments were replaced with similar sounds from the same group, the human ear could hardly detect any difference, with the naturalness score reaching 4.09. Additionally, as an optimization solution for the "inference phase," PCG does not require retraining existing models, and it only requires an additional 37MB of memory, paving the way for the future popularization of high-quality, low-latency AI voice services on various mobile terminals.
Key points:
🚀 Significant speed improvement: By introducing PCG technology, AI speech generation speed increased by about 40%, effectively solving the latency issue in text-to-speech technology.
👂 Stable audio quality: Replacing "exact matching" with "range verification" greatly improves efficiency while keeping the naturalness and speaker similarity of the audio almost intact.
🛠️ Low cost and easy deployment: This solution does not require retraining the model and only requires a minimal additional memory overhead, allowing direct application and optimization of existing AI voice inference systems.
